Field Experiments I
POLSCI 4SS3
Winter 2024
Last time
Hypothesis testing as a standard to evaluate if experiments suggest that policies works
We don’t know how things look when the policy work, but we can contrast with the hypothetical world in which it doesn’t work
We can only do this if we know how the experiment was conducted!
Lab: Statistical power as a diagnosand to determine if an experiment is well equipped to detect “true” effects
Today: Start applying these concepts to field experiments
Field Experiments
Field: Interventions in real world settings
(vs. surveys, laboratory)
Experiment: Randomization determines assignment of units to conditions
AKA randomized controlled trials in policy or A/B testing in industry
Core idea: Randomization allows us to produce credible evidence on whether something works
In practice: A lot of implementation details and research design choices to navigate
Examples
Banerjee et al (2021): TUP program
Poverty trap: Most programs to help the poor improve living conditions in the short term, but revert afterwards
Solution: Multidimensional “big push” to overcome poverty traps
Evaluate long-term effect on poorest villages in West Bengal, India
Household eligibility criteria
. . .
- Able bodied female member
(why?)
- No credit access
. . .
AND at least three out of
. . .
- Below 0.2 acres of land
(about 2 basketball courts)
- No productive assets
- No able-bodied male member
- Kids who work instead of going to school
- No formal source of income
Data strategy
Sample: 978 eligible households
514 assigned to treatment
266 accepted treatment
. . .
What was the treatment?
Program
Choose a productive asset
(82% chose livestock)
Weekly consumption support for 30–40 weeks
Access to savings
Weekly visits from program staff over a span of 18 months
. . .
Why would someone reject this?
Answer strategy
Track economic and health outcomes after 18 months, 3, 7, and 10 years
Of all household members
Focus on average treatment effect among the treated
(more in the lab)
Findings
Why does this work?
Time | Livestock | Micro-enterprise | Self-employment | Remittances |
---|---|---|---|---|
18 months | 10.26 | 7.93 | 18.67 | 0.00 |
3 years | 7.68 | 25.12 | 31.06 | 3.70 |
7 years | 27.26 | 67.59 | 108.36 | 8.87 |
10 years | 16.71 | 36.82 | 93.87 | 19.06 |
. . .
- Takeaway: Big push works because it helps people diversify their income sources over time
Full results on Table 3 of the reading
Pennycook et al (2021): Shifting attention to accuracy can reduce misinformation online
- Why do people share fake news in social media?
Three explanations:
- Confusion about accuracy
- Partisanship \(>\) accuracy
- Inattention to accuracy
Study 7: Application to Twitter
Studies 1-7 were all survey experiments
Study 7 deploys intervention on Twitter to see if priming accuracy works
N = 5,739 users who previously shared news from untrustworthy sources
Treatment: Send a DM asking to evaluate accuracy of news article
Challenge to data strategy
Can only send DM to someone who follows you
Need to create bot accounts and hope for follow-backs
Identify those who retweet fake news
Limit 20 DMs per account per day
3 waves with many 24-hour blocks in each
Stepped-wedge design
AKA staggered adoption design
Findings
The size of each dot represents the proportion of pre-treatment posts from that outlet
Next Week
More field experiments
Focus on: Sections 1-3 of Diaz and Rossitter (2023) only
Break time!